Overview
- Moonshot AI introduced Kimi K2.6 as an open-source coding model built to run multi-step tasks without supervision.
- The company describes a mixture‑of‑experts design with about 1 trillion total parameters and 32 billion active at inference.
- Vendor-run demos claim a 13-hour autonomous rewrite of the exchange-core matching engine with large throughput gains, though no independent audit is available.
- Moonshot also showcases persistent agents and large “swarms” that coordinate many sub-agents, plus a run that deployed Qwen3.5 locally and sped up a Zig-based inference loop.
- If the claims hold, developers could offload long refactors to agents, but teams would need stronger tests to catch hidden bugs and verify real-world correctness.